SIG-FPT 2026-09-18

Session Recording Summary · 51m 37s · Full notes ↗

Reading: Two papers on securing/robustifying stigmergic swarm systems. Neither is named with a full citation in the transcript, but they are clearly identified by content: - **Paper 1:** A study of ant-colony foraging under "misleading pheromone" attacks, introducing "smart pheromones" — richer signals that ID individual ants and their travel times, then use clustering (DBSCAN is mentioned) to identify and discard suspicious trails. Relies on a centralized/trusted nest authority. - **Paper 2:** The "TERMES"-style approach (referred to as "Termi's") — a compiler-based method that calculates backwards from a target structure how a decentralized stigmergic swarm should build it, then probabilistically decays error-prone points to keep states reversible.

The session (apparently the last reading discussion of a summer series) compared two contrasting strategies for making stigmergic systems robust against corruption: enforcing correctness by enriching signals with metadata/identity vs. ensuring errors remain reversible via probabilistic decay. Participants went around the call sharing reactions, converging on a preference for the decay/reversibility approach and a shared discomfort with adding centralized authority or identity systems. The back half of the session shifted to project logistics for an upcoming hands-on workshop (a swarm/ant simulator), covering a pending PR, traceability/replay features, and playtesting.

• **Paper 1** breaks the indistinguishability of the basic stigmergic setup by adding per-ant identity and travel-time metadata; slower/divergent ants accumulate "suspicion" and are eventually excluded. Uses clustering to separate legitimate food-source connections from bad ones.
• **Paper 2** is less centralized: given a known target structure, it identifies points where an error would be catastrophic and then **probabilistically decays** those points (uniformly, concentrated at critical points, or time-varying with less decay later). Patrick called this the more interesting direction than the DBSCAN approach, though it risks **deleting too much** and hindering the structure.
• **Paper 1 = enforce correctness** by infusing knowledge/metadata into signals (loaded signals).
• **Paper 2 = don't pursue correctness; add mortality** — make states non-permanent and let the system heal, relying on the "if-mortality" that correct dynamics will drive correct reconstruction. Closer to true stigmergy given the locality-of-knowledge constraint (agents should only know local state; intelligence should be emergent, not infused).
• **Coordinate reversal for AI:** the readings assume the swarm is the blue team defending against a centralized red attacker; AI flips this — the defended entity is often centralized and the **attackers are swarms**. He mapped the four quadrants (centralized vs. centralized = traditional security; centralized blue vs. swarm red; swarm blue vs. centralized red = many of these papers) and argued the truly interesting case is **swarm vs. swarm**.

Questions & Disagreements: - **Ergod's central tension:** At what point does defending a stigmergic system require reconstructing the very coordination machinery stigmergy is meant to avoid (i.e., injecting external knowledge or extra defensive constructs)? Can a protocol allow agents to decay/corroborate/route-around traces — doing *more* than a single agent's local remit — without breaking the pure-stigmergic character? -

Participants: Patrick Nast, Ergod, ComplexSy, Venkatesh Rao, Wip, Giovanni Merlino, Dan Schmidt